[Quant][Perf] Use moe_wna16 kernel by default for MoEs with many experts (#13236)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-02-14 12:53:42 -08:00
committed by GitHub
parent c9e2d644e7
commit 5e5c8e091e
4 changed files with 39 additions and 26 deletions
@@ -9,13 +9,8 @@ from vllm.model_executor.layers.fused_moe.layer import (
FusedMoE, FusedMoEMethodBase, FusedMoeWeightScaleSupported)
from vllm.model_executor.layers.linear import (LinearBase,
UnquantizedLinearMethod)
from vllm.model_executor.layers.quantization.awq import AWQConfig
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.layers.quantization.gptq import GPTQConfig
from vllm.model_executor.layers.quantization.gptq_marlin import (
GPTQMarlinConfig)
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
check_marlin_supports_layer)
from vllm.model_executor.utils import set_weight_attrs
@@ -37,6 +32,12 @@ class MoeWNA16Config(QuantizationConfig):
self.linear_quant_method = linear_quant_method
self.full_config = full_config
self.use_marlin = False
# Avoid circular import
from vllm.model_executor.layers.quantization.awq import AWQConfig
from vllm.model_executor.layers.quantization.awq_marlin import (
AWQMarlinConfig)
from vllm.model_executor.layers.quantization.gptq_marlin import (
GPTQMarlinConfig)
if self.linear_quant_method == "gptq":
self.use_marlin = GPTQMarlinConfig.is_gptq_marlin_compatible(
full_config)
@@ -115,6 +116,8 @@ class MoeWNA16Config(QuantizationConfig):
capability_tuple = current_platform.get_device_capability()
device_capability = (-1 if capability_tuple is None else
capability_tuple.to_int())
# Avoid circular import
from vllm.model_executor.layers.quantization.awq import AWQConfig
awq_min_capability = AWQConfig.get_min_capability()
gptq_compatible = quant_method == "gptq" and \
@@ -129,6 +132,13 @@ class MoeWNA16Config(QuantizationConfig):
if is_layer_skipped_quant(prefix, self.modules_to_not_convert):
return UnquantizedLinearMethod()
elif isinstance(layer, LinearBase):
# Avoid circular import
from vllm.model_executor.layers.quantization.awq import AWQConfig
from vllm.model_executor.layers.quantization.awq_marlin import (
AWQMarlinConfig)
from vllm.model_executor.layers.quantization.gptq import GPTQConfig
from vllm.model_executor.layers.quantization.gptq_marlin import (
GPTQMarlinConfig)
if self.linear_quant_method == "gptq":
if self.use_marlin:
return GPTQMarlinConfig.from_config(